Search Papers | Poster Sessions | All Posters
Poster C105 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Sparse predictive-coding networks account for Bayesian and ‘anti-Bayesian’ effects in human orientation perception
Stefan Brugger1,2, Christoph Teufel1; 1Cardiff University, 2University of Bristol
Natural scenes are dominated by horizontal and vertical local orientations. Bayesian models of vision therefore suggest that the visual system implements a prior biasing orientation perception towards cardinal orientations. The existing evidence, however, suggests that this view may be too simplistic: while neuroimaging studies report neural representations biased towards cardinal orientations, psychophysical work suggests that perceived orientation is biased away from cardinal orientations. Here, we reconcile these findings using neural-network modelling combined with psychophysical testing. We implemented a sparse predictive-coding network as a biologically-plausible model of perception and learning in the visual system. Following training on natural scenes, orientation processing was tested with orientated gratings of varying signal-to-noise ratio. In line with previous work, the network developed orientation-tuned receptive fields. Anisotropy emerged spontaneously, with greater preponderance of units tuned to cardinal than oblique orientations. This non-homogeneity acted as a structural constraint, reproducing the oblique effect seen in human vision, as well as generating attractive biases towards cardinal orientations in neural representations. Importantly, due to lateral inhibition, biases increased with stimulus signal-to-noise ratio. Consequently, in simulated psychophysical experiments, the network reproduced the pattern of apparent repulsive biases seen in human observers. These results are able to reconcile apparently contradictory findings in human psychophysics and visual neuroscience.
Keywords: predictive coding orientation perception sparse coding vision